In the dynamic field of residential mortgage lending, the accurate estimation of loss given default (LGD) is crucial for effective risk management and model validation. This study investigates LGD using machine learning techniques on a data set of conforming loans from government-sponsored enterprises in the United States between January 2012 and March 2018. The research objectives were threefold: to evaluate the predictive capabilities of various machine learning models for LGD that incorporate a comprehensive set of risk factors; to compare LGD across lender types; and to identify the dominant variables influencing LGD outcomes. Ensemble tree-based methods (specifically the CatBoost, the Gradient-Boosting Regressor and Light Gradient-Boosting Machine regressor algorithms) demonstrated superior accuracy, which was further enhanced through hyperparameter tuning. These models highlighted the importance of loan-specific details and broader economic indicators in LGD estimation. The study found that, unlike banks and nonbanks, fintech firms presented distinct LGD patterns, suggesting diverse risk assessment approaches. Key variables impacting LGD predictions included the mark-to-market loan-to-value ratio, the unpaid balance at default, the liquidation timeline and the mortgage servicer as well as regional economic factors. This paper contributes to the methodological advancement in LGD prediction at the origination stage, which is crucial for proactively managing loan portfolios. It contributes to advancing methodological approaches in LGD prediction and risk model validation, offering valuable insights for lenders and policy makers in refining risk frameworks within the evolving mortgage market.